#reinforcement learning LLM
870 related articles
·2 min
Linguists Enter the AI Arena: The Human Truth Behind LLM Evaluation
A $40-50/hr linguistics expert job reveals the truth behind AI training: why LLM evaluation needs native-speaker experts and how RLHF human feedback determines model quality ceilings.
Read more →
·3 min
The State of Reinforcement Learning in Montezuma's Revenge: From Exploration Challenges to General Intelligence
In-depth analysis of Montezuma's Revenge in RL research: reviewing Go-Explore and RND breakthroughs, and the shift toward sample efficiency and generalist agents.
Read more →
·3 min
SJTU's ARIS Framework: Making Research Agents Reliably Conduct Autonomous Science
Shanghai Jiao Tong University releases ARIS framework for reliable end-to-end research automation. Self-review loops, score thresholds, and human-in-the-loop design solve AI agent drift problems.
Read more →
·4 min
OpenAI Chief Research Officer Mark Chen: Within Three Years, AI Will Independently Conduct End-to-End Scientific Research
OpenAI CRO Mark Chen shares frontier AI research insights: RL boundaries, why Scaling Laws aren't dead, the o1 reasoning model's origin story, and the bold three-year goal of AI conducting end-to-end scientific research independently.
Read more →
·4 min
Understanding RAG from Scratch: Give a Large Model Context and It Can Answer Anything
What is RAG (Retrieval-Augmented Generation)? This article explains RAG core concepts with simple analogies, analyzes three LLM pain points, and details RAG's working mechanism and future trends.
Read more →
·3 min
How to Do Research in the AI Era: Avoid Outsourcing Your Thinking to Machines
When AI can instantly read papers and generate code, how can researchers avoid cognitive atrophy? This article explores the traps of AI-assisted research and offers practical advice for rebuilding methodology.
Read more →
·4 min
Advancing Your Self-Taught Machine Learning Journey: Three Paths Forward After Completing Foundational Projects
After completing MNIST implementation and paper reproduction, how should self-taught ML learners advance? This article outlines three paths: computer vision, NLP, and math foundations.
Read more →
·3 min
LLM Learning Roadmap: A Three-Phase Practical Guide
A systematic LLM learning roadmap: from Python basics to LangChain & LlamaIndex frameworks, RAG, Agent, and fine-tuning core skills, plus hands-on projects to master LLM app development in 3 months.
Read more →
·3 min
What Are AI Agents? A Deep Dive into the Three Core Components: Perception, Decision, and Action
A deep dive into AI Agents: their definition and three core components—Perception, Decision, and Action. Learn what distinguishes real AI agents from chatbots and automation scripts.
Read more →
·3 min
Scaling Laws for Pre-training and RL: Chess Experiments Reveal Post-Training Compute Allocation Principles
Chess experiments systematically study compute allocation across pre-training, SFT, and RL, revealing that pre-training sets the downstream ceiling and RL mainly boosts pass@1 reliability, not exploration breadth.
Read more →
·2 min
Reasoning Chain Theft Attacks on Closed-Source LLMs: Security Risks of API Side Channels
Are hidden reasoning chains in closed-source LLMs truly secure? Research shows attackers can reconstruct full thought chains via API side-channel signals, threatening trade secrets and IP.
Read more →
·2 min
AI Store Management Experiment: Friendly but Not Smart Enough — Revealing the True Capability Boundaries of AI Agents
A new study had AI independently run a store, revealing that AI shopkeepers are friendly but make poor business decisions. Analysis of AI Agent real-world capability limits.
Read more →
·3 min
Self-Interrogation: A Novel Approach to Reverse Engineering DeepSeek by Interviewing the AI
Exploring an innovative approach to reverse engineering DeepSeek by directly interviewing the AI assistant, analyzing system prompt leakage, hallucination issues in model self-descriptions, and implications for AI transparency and prompt injection security.
Read more →
·3 min
The Vision of AI for All: The Path to a Positive AI Future Shared by Everyone
Exploring how AI can transform from an exclusive tool of tech giants into a shared capability for all humanity. Analyzing key paths and challenges through open source, education, and governance.
Read more →
·3 min
Why Does Gemini Keep Getting Things Wrong? A Deep Dive into AI Hallucinations and How to Deal with Them
Deep analysis of why Google Gemini and other LLMs frequently produce errors, explaining the technical mechanisms behind AI hallucinations and offering practical prompting tips for better AI usage.
Read more →
·2 min
Which Programming Language Is Best for AI Coding Assistants? The Battle Between Type Systems and Training Data
Exploring language choice in the AI coding assistant era: statically typed languages like TypeScript and Rust enable AI self-correction via compiler feedback, while Python leads with massive training data.
Read more →
·2 min
AI Overcompensation: Why Models Can't Just Say a Single Word
Analyzing why AI models can't just say a single word when asked — exploring the technical causes behind overcompensation, from RLHF training bias to instruction-following limitations.
Read more →
·3 min
Claude vs GPT Knowledge Cutoff Dates Explained: How to Determine AI Knowledge Boundaries
A deep dive into knowledge cutoff dates for LLMs like Claude and GPT, covering pre-training data endpoints, how to verify AI knowledge boundaries, and how RAG overcomes time limitations.
Read more →
·3 min
Does Saying Please and Thank You to ChatGPT Actually Matter? A Deep Dive into AI Politeness
A viral tweet about a wife worried she's annoying the people behind ChatGPT. Exploring human instincts to anthropomorphize AI, the real value of politeness toward AI, and maintaining humanity in human-machine interaction.
Read more →
·3 min
AI Is Not a Genie in a Lamp — It's a Meeseeks: A Risk Metaphor for Goal-Driven AI
Using Meeseeks from Rick and Morty to analogize AI safety issues — more precisely revealing intrinsic motivation risks, instrumental convergence, and corrigibility challenges in goal-driven agents.
Read more →